Anthropic raised $65B and shipped Opus 4.8, Microsoft got back in the coding race, and labs collectively sat down to design their own chips.
🤖 AI
Anthropic updated its flagship
Claude Opus 4.8 is out: better effort control, dynamic workflows in Claude Code, and a cheaper fast mode. Simon Willison (developer, creator of Datasette) called the update:
> "Modest but noticeable" — Simon Willison
MiniMax is teasing a leaner M3
In the M2 technical report, the company showed off sparse attention — it speeds up long-context responses by up to 15.6×, which finally makes long context affordable for agents.
NVIDIA put agents inside a "world model"
γ-World manages independent agents in real time and transfers skills from two players to four without any fine-tuning.
Open models are falling behind again
Their gap with closed models has been estimated at 4–6 months: it nearly closed during the DeepSeek R1 era, and now it's widening again.
💻 Software
Claude Code rewrote Bun in Rust by itself
With the new Dynamic Workflows, Jarred Sumner (creator of the Bun JavaScript runtime) migrated Bun from Zig to Rust — 750,000 lines in 11 days with 99.8% tests passing: the model breaks down tasks and runs agents in parallel.
Cursor showed how developers actually code
The habits report found that models pull in more codebase context — and that alone makes work cheaper and boosts the rate of accepted edits.
Microsoft gets back in the coding race
The company is building its own code-generation model — a response to competitors it had been feeding money to itself. OpenAI Codex and Claude Code already own this market.
⚙️ Robots and hardware
ByteDance got tired of waiting for chips
The company is going to make its own processors: fed up with waiting months for deliveries, it's looking for partners for AI hardware.
Mistral is going into silicon
Arthur Mensch (Mistral CEO) wants his own chips — to control the infrastructure and cut data center costs in Europe.
💰 Money
Anthropic raised $65B
Series H round at a $965B valuation — enterprise demand, $47B annual run-rate, and a rumored first profitable quarter on the horizon.
AI is actually paying off
Willison unpacked the paradox: token bills terrify businesses not because AI doesn't work, but because it does — he personally racked up $2,180 in a month on a $200 subscription. More on how LLM prices are falling.
Musk downplayed the Anthropic deal
Publicly he called SpaceX's compute contract non-binding, but in the S-1 filing the same deal is described as a three-year commitment.
🎭 Off the record
The Vatican weighed in on AI
Pope Leo XIV released the first encyclical on artificial intelligence — Willison put together a summary of the key theses.